A novel approach combining temporal and spectral features of Arabic online handwriting for Parkinson's disease prediction.

Background: Parkinson's disease (PD) affects millions of people worldwide, and it is predicted that this pathology will gravely increase in the next few years. Unfortunately, there's currently no cure for this disease, indeed an early diagnosis of Parkinson's disease can help to better manage its symptoms and its evolution. One of the most frequent abilities and usually also the first manifestation of Parkinson's disease is alteration of handwriting.

New Method: We propose a novel method to detect Parkinson's disease, based on the segmentation of the online handwritten text into lines. Indeed, we propose to compare Parkinson's disease patients and healthy controls, based on the full dynamics of new temporal and spectral features. Three classifiers were used, K-Nearest Neighbors, Support Vector Machine and Decision Trees. The performances of these three classifiers were estimated using a stratified nested 10 cross-validation. All the models in this study have been evaluated using classification accuracy, balanced accuracy, sensitivity, specificity, F-Score and Matthews Correlation Coefficient.

Results: An accuracy of 92.86 % was obtained with Decision Trees classifier in the last line. The new categories of spectral and temporal features gave the best classification performances in comparison to the basic statistical features.

Comparison With Existing Methods: Previous studies have only focused on words or sentences. This is the first study to deal with the analysis of a text composed of several lines.

Conclusion: The last line discriminates at best between Parkinson's disease patients and healthy controls. This obtained result has further strengthened our hypothesis concerning the fatigue occurring while writing in PD patients.